Migrating to the new server part 1: base system and the web server

Last update:

Tags: servers, self-hosted

My current VPS is running CentOS 6, and, frankly, it long started showing its age. Not very surprising if we remember that it was released in 2011. Even with all efforts from EPEL, Remi, and Software Collections maintainers, running new applications on new OS versions is obviously easier than on old verions. That's why I decided to migrate to Fedora, and it was much easier than I thought.

Choosing the platform

Despite the amount of time I spend messing up with Debian in the VyOS project, ultimately I’m a RedHat guy. Or at least an RPM-based distro guy. I was a fan of the original RedHat back in early 2000’s when I was a high school student, then used OpenSuSE for a while, then switched my desktop to Debian solely because of my involvement with Vyatta, but eventually switched to Fedora a few years ago and moved my VyOS development to dedicated Debian machines. Sorry Debian fans, nothing against it, it’s just not my cup of tea.

An obvious choice would be CentOS 7, but… It’s already a few years old, in two years it will be switched from full support to life support, it still only has Python 2.7 in its main repos, and due to being systemd-based it requires as much migration as everything that came after it. The worst part, however, is that you cannot upgrade CentOS to the next version without reinstalling the machine. My VPS went through miration from CentOS 5 to 6. This migration is unavoidable, but I would rather not repeat it any soon afterwards.

What I liked about CentOS is that its long release cycle doesn’t force you to migrate to a new version any soon. However, I would be fine with just stable enough for my needs (doesn’t break on updates) rather than ABI-stable for years. I don’t run enterprise applications or anything after all. However, going for a rolling release distro that would solve that problem by continuous upgrades, doesn’t feel safe, as they do break once in a while.

In the few years of using Fedora, I’ve updated it with dnf system-upgrade every time, last time even skipping a version (from 25 to 27), and it never broke. So I thought this is a good compromise that is already confirmed by my experience.

Ever since I got my first VPS in 2010, it’s been something of a tradition that my friends who run small hosting companies host it for me in exchange for help with network equipment configuration. So, Alexander Norman who hosts my current server generously setup a second machine for me and installed Fedora 27 server on it as per my request, and the fun has begun.

Setup choices

If this is a big migration anyway, it’s also a good time to rethink the choice of software and setup. So far I decided to stick with the same software for the most part, but I looked for alternatives.

I looked into Docker or plain LXC, but ultimately decided against it: not quite the use case for them, and since the machine is self-contained and relatively limited in storage space, the added complexity and storage overheard are probably not worth it.

The most important decision was to use SELinux in enforcing mode. Containers only isolate applications from one another, but offer no additional hardening for applications themselves, so something would have to be done about it anyway. Besides, I just wanted to learn more about SELinux. The operations side of my career has been mostly focused on networks and virtualization for quite a long time, and I got a bit out of touch with the rest of the system administration world.

I also decided to enable HTTPS for all websites, because why not, everyone is doing it. The stuff I write doesn’t pose privacy concerns for anyone (unless being a nerd is illegal or immoral anywhere), but it’s still a good thing to have.

My email system is mostly in sync with the latest security developments (DKIM etc.) so it likely won’t need many changes. I’m going to try to setup DNSSEC however.

I’ve also decided to keep the original hostname and the login banner associated with it:


-- When you're on a bike, the ocean is a lot closer than you think.
-- The autumn salt wind went right through to the back of my nose.
-- And maybe it's because, like Haruko said, my head was empty.

It comes from FLCL, which you should watch/read if you enjoy surreal coming of age stories. Old nerds never grow up, you know.

But, in this part I’ll describe the base system and web server setup as I go in the tradition of gonzo journalism, and leave the other parts for later.

Filesystem layout fixup

My friend installed Fedora with some of the default disk layout options, and it created a separate logical volume for /home. I assumed that the default would be everything in one partition, and didn’t tell him how I want it setup, so I can’t blame him.

Still since a lot of data such as email, web content, and databases will live in /var, while /home will be relatively empty, this setup just makes no sense for me. Good the default option used LVM. I’ve temporarily allowed root login in /etc/ssh/sshd_config to login without opening any files in /home. Them, assuming that /boot is on /dev/sda1 and the physical volume is on /dev/sda2, it was just a matter of:

# mv /home/dmbaturin /tmp
# umount /home
# lvremove /dev/fedora/home
# lvextend /dev/fedora/root /dev/sda2
# resize2fs /dev/mapper/fedora-root
# mv /tmp/dmbaturin /home

Since lvextend automatically fills all free space on the physical volume if you specify the device it’s on instead of size, and resize2fs fill all free space on the volume if no size argument is given, I didn’t even need to do any size calculations in this case.

Compared to what we had to endure in the days of plain partitions and filesystems that could only be resized when not mounted, LVM and ext4 make this part anticlimactic.


I can see why one may want firewalld on their desktop, but on servers, I think it just has no place there. It doesn’t add any value if the rules are all predefined, but does add complexity. So, normal iptables then, but the init scripts for loading rules on boot are no longer there by default.

Luckily, the package that has systemd unit files is right there in the repos:

# iptables-services
# systemctl enable iptables.service
# systemctl enable ip6tables.service

Just put your rules in /etc/sysconfig/iptables and /etc/sysconfig/ip6tables respectively. You can use iptables-save to save them in the loadable format.

My rules are pretty straightforward, for services I will configure first I need something like:

[root@haruko ~]# cat /etc/sysconfig/iptables

# Keep state

-A INPUT -m state --state NEW -m tcp -p tcp --dport 22 -j ACCEPT -m comment --comment "SSH"

-A INPUT -m state --state NEW -m tcp -p tcp --dport 80   -j ACCEPT -m comment --comment "HTTP"
-A INPUT -m state --state NEW -m tcp -p tcp --dport 443  -j ACCEPT -m comment --comment "HTTPS"

-A INPUT -p icmp -j ACCEPT

# Local traffic
-A INPUT -i lo -j ACCEPT

# Default actions
-A INPUT -j REJECT --reject-with icmp-host-prohibited
-A FORWARD -j REJECT --reject-with icmp-host-prohibited

It could benefit from ipset perhaps, but it can wait.

What about nftables?

I want to love nftables, but at this point, I just can’t. Yes, it’s a promising tool, and I’m sure the kernel part is good already, but when the most recent release of the userspace tool cannot display rule counters (even though it can configure rules to use counters), that’s… a tough sell at the very least. Also, its man page is one of the worst I’ve seen: the man pages of iproute2 give a complete and correct command argument grammar in BNF at least.

I’ll be watching the situation, but for now, I’m sticking with what actually works.

Apache HTTPD and letsencrypt

My original plan was to prepare configs for all websites complete with HTTPS on the new server first and then move the data at once. My old server was using StartSSL certificates originally. StartSSL was fun while it lasted I have to admit. Now it’s gone, so I needed an alternative, and Let’s Encrypt seems like an obvious choice by now.

For better or worse, my original plan failed. Doing HTTP challenges by hand is a royal headache, and the certonly option in general is not so nice either. At first I was sceptical and started thinking of getting certificates from a provider like Gandi for $50/year. But I decided to try moving what’s easy to move and give the automated workflow with certbot Apache HTTP plugin first, and I wasn’t disappointed.

On the old server, HTTPS was only enabled for applications that really need it, while public websites were HTTP only. All sensitive applications were under a single SSL virtual host. It turned out my setup was based on an assumption that is no longer true: that you cannot have name-based virtual hosts with SSL and use different certificates for them all.

Sure, it relies on a TLS extension, but as this Apache wiki page suggests, clients that do not support it have long reached EOS, and if they are not completely extinct by now, it’s their problem, not mine, so there’s no reason to care about them. Really, who in their right mind is still using Internet Explorer 7 and Windows XP, or pre-2.0 Firefox?

My Apache HTTPD config layout is a bit peculiar, I like to keep all virtual hosts configuration files in /etc/httpd/vhosts.d/*.conf and include them with Include vhosts.d/*.conf from /etc/httpd/config/httpd.conf. I was surprised to find out that certbot, when used as intended, picked that up and correctly created additional config files for HTTPS virtual hosts under vhosts.d. So I’ve just ran certbot --apache, picked all websites, and it did its trick rather nicely. Didn’t even override anything in the original configs, just added a redirect to HTTPS as I’ve asked.

I’m still to migrate some other stuff and see if adding HTTPS for a new one will go just as smoothly, but I’m pretty confident that it will.

LAMP and SELinux

Remember that I decided to keep SELinux on this time? Installing LAMP is trivial, but making it pass audit checks took me a bit more time.

Before I tell you what’ve done, save this command somewhere, it’s going to help you a lot.

ausearch -m avc -ts today | audit2why

It will always tell you some solution for making the audit error go away. It will not always be the best solution, however, but rather a universal one. If you are into SELinux, you may want to use something more granular instead, for example, correctly set file context instead of disabling checking it.

Matomo (formerly Piwik) was the first web application I decided to install, so I’ll use it as an example.

Allowing Apache to read files

So, as usual I copied the Matomo files to /var/www/vhosts/matomo.baturin.org and did chown -R apache:apache on it. And right at the stage when it checks if it can write to tmp/ directory, I got this:

In logs:
(13)Permission denied: [client ...:48310] AH00035: access to /index.php denied 
(filesystem path '/var/www/vhosts/.../index.php')
because search permissions are missing on a component of the path

In ausearch:

type=AVC msg=audit(1520619405.318:916444): avc:  denied  { getattr } for  pid=3961 comm="httpd"
 path="/var/www/vhosts/.../index.php" dev="dm-0"
ino=1575637 scontext=system_u:system_r:httpd_t:s0 
tcontext=unconfined_u:object_r:admin_home_t:s0 tclass=file permissive=0

Was caused by:
  Missing type enforcement (TE) allow rule.

The solution is to set the right context for those files so that Apache HTTP is allowed to read them.

chcon -R -t httpd_sys_content_t /var/www/vhosts/matomo.baturin.org/

Without looking further you can usually find out what the context should be by looking at the directories created by packages with ls -alZ.

Allowing Apache HTTPD to write to files

Now the Matomo installer page loads, but complains that tmp/ directory is not writable. In ausearch we see this:

type=AVC msg=audit(1520627385.839:1094): avc:  denied  { write } for  pid=3829 comm="php-fpm" name="matomo.baturin.org"
dev="dm-0" ino=1704836 scontext=system_u:system_r:httpd_t:s0
tcontext=unconfined_u:object_r:httpd_sys_content_t:s0 tclass=dir permissive=0

The solution audit2why suggests is to set httpd_unified SELinux boolean to true. That would allow HTTPD to write to any files with httpd_sys_content_t context though. Is it really what we want? I thought I like the idea of separating application files from writeable data and instead set the context of those directories to one that allows writing: httpd_sys_rw_content_t.

chcon -t httpd_sys_rw_content_t /var/www/vhosts/matomo.baturin.org/tmp/
chcon -t httpd_sys_rw_content_t /var/www/vhosts/matomo.baturin.org/config/

Allowing database connections

Now system checks pass, but Matomo installer complains that it cannot connect to MySQL. There are two SELinux booleans that would allow it: httpd_can_network_connect and httpd_can_network_connect_db. The latter only allows connections to ports and sockets of well known database servers such as PostgreSQL, MySQL, and Oracle. Since I don’t have web applications that would need generic network connections, I went for the second, more granular option:

setsebool -P httpd_can_network_connect_db=1

If you are wondering what exactly this or that boolean does, you can view the policy associated with it with this command:

sesearch -A -b httpd_can_network_connect_db


Web server setup works, and that’s a good progress. Next step is the email server, but that’s for a later post.